DocExtr Changelog

0.1.1 2026-10-11

Extraction results are now available for four purposes: AI agents, embeddings, search, and plain text. Output formats, some commands, and the Python API are not compatible with 0.1.0.

Added

• Image OCR — reads text in images embedded in documents and in standalone image files. Turn it on with --ocr true --ocr-models DIR; the default is false. The CLI and Python wheel include the OCR engine
• OCR targets — pages or slides with no body text, or unit numbers you specify. --ocr-pdf-render true renders the selected PDF pages in full for recognition
• OCR text from note images stays with the notes, separate from the body, and the recognized text is saved as a .txt file next to each exported image
• agent results separate per-unit body text (body_text, body_markdown) from notes. Parsed PDF pages and PPTX slides keep their structure even when the body is empty, and PPTX design is preserved

Changed

• --targets selects only the results you need (all four by default): agent gives Markdown with repeated tables of contents and headers cleaned up, plus document structure and images; text gives plain text and Markdown faithful to the source; embedding gives chunks sized by estimated token count; and search gives search documents sized by character count. Chunk size and overlap are set separately for embeddings and search
• extract and batch write to per-document directories, JSONL shards, or a database. batch --layout document mirrors the input folder structure with files for each document, and stores archive contents under a directory named after the archive with a .vfs/ suffix
• File output consists of a manifest with the document status plus one set of files per requested purpose. Database output is split into a document table and embedding and search chunk tables
• Python extraction functions and Engine return results by purpose, and export_file writes the same file layout

Breaking changes

• CLI run is renamed batch, with no alias for the old name
• document.json, chunks.jsonl, and the 11-field JSONL and CSV output are replaced by the manifest and per-purpose outputs. --chunks, --chunk-*, --format, --profile, and --no-markdown are removed; get chunks with --targets embedding or search and that purpose's size and overlap options
• The single database table from 0.1.0 is not converted automatically. Load the data again under a new table name
• In Python, purpose-based result dicts replace the Document dict and extract_chunks. Embedding chunks are in embedding.chunks of the targets=["embedding"] result
• OCR options accept only true and false. off, auto, force, and --ocr-max-pages are removed. To read whole PDF pages, use --ocr true --ocr-pdf-render true
• In batch, --cache-dir and --max-cache-bytes replace --cache-bytes and --cache-entries, and --no-cache turns the cache off
• extract leaves a manifest and exits with code 3 when extraction fails or is partial

Downloads

• CLI — Linux x86_64 and Windows x64, with OCR and database output
• Python wheel — Linux x86_64 (manylinux_2_28), CPython 3.9–3.14. Includes the OCR engine; database output and job state are not included
• OCR model (PaddleOCR-VL 1.6 Q8) — not included in the CLI or wheel; install it separately
• Licenses — the CLI and Python wheel are MIT. Third-party notices and the original source of MPL-2.0 components (docextr-0.1.1-third-party-sources.zip) are provided alongside

Known limitations

• 7z, RAR, HWP 4.x, and encrypted documents are not supported
• The batch memory budget does not enforce a cap on total memory use, including parsers, OCR, and database drivers
• No macOS executables are provided. The Windows executable has not been verified on Windows itself; only text extraction was checked under Wine (OCR and database output unchecked)

0.1.0 2026-10-01

First public release.

Added

• CLI docextr run — batch-processes files, directories, and the archives inside them into JSONL or CSV files or database tables. Errors are recorded in the record for that file and the rest of the batch continues
• CLI docextr extract — exports one document to a new directory as document.json, document.md, and assets/. With --chunks it also writes RAG chunks to chunks.jsonl
• Formats — HWP/HWPX, text PDF, DOC/DOCX, PPTX, XLS/XLSX/XLSB/ODS, TXT, Markdown, HTML, and source and configuration files
• Archives — ZIP (encrypted ZIP with --archive-password-file), TAR, TAR.GZ, and TAR.BZ2 processed recursively, with nesting depth and extracted size limits
• Scanned PDF OCR — runs with --ocr auto or force and a separately supplied model directory. The default is off
• The 11 output record fields (source_path, plain_text, markdown, blake3, file_type, file_size, error, title, author, created, word_count) are shared by JSONL, CSV, and database output
• Database output --output-db — loads into SQLite, PostgreSQL, MySQL, and MariaDB in append, upsert, or replace mode
• Additional run options — structured output --profile structured, result cache --cache-bytes, saving and resuming job state --state and --resume, and a failure limit --max-failures
• Python package docextr (CPython 3.9–3.14) — extract_file, extract_bytes, extract_files, export_file, extract_chunks, and an Engine class with a cache

Downloads

• CLI — Linux x86_64 and Windows x64, with OCR and database output. The Windows executable has been run only under Wine
• Python wheel — Linux x86_64 (manylinux_2_28), CPython 3.9–3.14. OCR and database output are not included
• Licenses — the CLI and Python wheel are MIT. Third-party notices and the original source of MPL-2.0 components (docextr-0.1.0-third-party-sources.zip) are provided alongside

Known limitations

• 7z, RAR, HWP 4.x, and encrypted documents are not supported
• The run memory budget does not enforce a cap on total memory use, including parsers, OCR, and database drivers
• No macOS executables are provided, and the Windows executable has not been run on Windows itself